Ressources : les données lues par le modèle

After tools comes the resource: read-only data your server hands to the model as context, like a file, a record, or a config value. 📄

A resource never changes the world. It only exposes information the model can read, the way a GET request fetches a page without editing anything.

Each resource is addressed by a URI, such as file:///notes.txt or config://app. The scheme and path tell clients exactly what they are asking for.

In Python you mark a function with @mcp.resource() and pass its URI. The function returns the content whenever a client reads that URI.

What your resource returns becomes the content the model reads. Plain strings work great for notes, settings, or any small document.

A resource also carries a MIME type like text/plain or application/json. It tells the client how to interpret the bytes it receives.

Resources are usually application-controlled. The host app or user chooses which ones to attach, rather than the model grabbing them on its own.

A client can ask your server to list its resources. That discovery step lets a host show the user which documents it can pull into context.

To fetch one, the client sends a read request with the URI. Your server runs the matching function and returns the current content.

Good resources are things you'd want the AI to see but not run: a README, a database row, log output, or the latest settings.

A simple test: if it does something, make it a tool; if it just provides information to read, make it a resource. Action versus context.

Let's confirm what a resource is for.

Great! A resource is read-only context with a URI and MIME type. Clients list and read them, and the model gains information without running anything. ✅